rl-CARLA

The basic idea is using Raw Image as state spaces to train DDPG Agent. The network architecture is quite simple, if you want to know more, you can check here. In order to evaluate the performance of the RL method, we first used supervised learning to train a network as baseline. Then we investigate the performance of RL methods (DDPG), both with and without pretraining.

Demo

imitation

DDPG

Result

              Town1(Train)                              Town2(Test)

Installation

Clone the git repository

$ git clone https://github.com/zhangfuyang/DDPG-CARLA.git

$ cd DDPG-CARLA & pip -r requirements.txt

(Use $DDPG_DIR as the root directory of the source)

Install Carla Simulator

1. Download the Carla you can just download the compiled version from here . Make sure the version of the simulator is 0.8.2(stable), I'm not sure if other development or old stable versions are compatible.

2. extract to the directory you want $ tar -xvf CARLA_0.8.2.tar.gz $CARLA_DIR

Training

1. First start the Carla server cd $CARLA_DIR ./CarlaUE4.sh -carla-server -benchmark -fps=10

2. Run ddpg_main.py cd $DDPG_DIR python ddpg_main.py

Test the model

1. Start the Carla server

2. python test_ddpg.py -model_path='models/'

Some Detail of implementation

1. Use pretrain Network https://github.com/carla-simulator/imitation-learning to extract feature. Then connected with actor and critic net.